Where will the AI ​​take people

Artificial intelligence is not interested in us humans

Prof. Dr. Jürgen Schmidhuber

Jürgen Schmidhuber is one of those AI experts who has a lot of confidence in intelligent systems. But although they will soon be far more intelligent than us, we need not fear them: they will not care about us.

EnlargeAI will soon overtake us - and leave us behind?

Jan-Bernd Meyer, editor-in-chief of Computerwoche, has Prof. Dr. Jürgen Schmidhuber interviewed about it.

In your lectures you will reflect, among other things, on how the most important events from a human perspective since the Big Bang have accelerated exponentially in a beautiful pattern. In this context, you also make predictions of where artificial intelligence (AI) will lead. Google is currently promoting the topic with the acquired British company Deepmind: In the game of Go, Google Deepmind's system "AlphaGo" beat the world's best human player. What is the relationship between Deepmind and your lab?

Schmidhuber: Deepmind's first doctors of AI once met in my research group, one became a co-founder, a first employee. More of my PhD students later joined Deepmind.

What do you think of AlphaGo?

Schmidhuber: I am of course happy about this great success. AlphaGo is based on more traditional methods, because in 1994 an initially stupid neural IBM program learned in a very similar way to become as good as the best backgammon player in the world by playing against itself. But there were a few new tricks, and most importantly, computers today are 10,000 times faster per euro. Today, games like backgammon, chess and go are in the hands of computers.

Do computers now do all of our games better than we do?

Schmidhuber: Just the board games. Physical games such as soccer are incredibly much more difficult because everything comes together there: rapid pattern recognition in the real world, which is much more complex than board games, fine motor coordination of complex movement sequences in partially observable surroundings, etc.

Please note: pattern recognition alone is generally much more difficult than chess. Since 1997 the world's best chess player is no longer human. But back then, computers were far inferior to any child in recognizing visual objects or speech. That has only recently changed with our neural networks - it was not until 2011 that our team achieved the first superhuman visual pattern recognition results in a competition in Silicon Valley.

In football, however, you have to be able to do a lot more than just pattern recognition - currently no robot can even come close to keeping up with human footballers. Although it won't stay that way in the long run.

So there will be AI systems in the near future that will outperform humans in every way?

Schmidhuber: I think so.

How will they work?

Schmidhuber: They will use our Artificial Feedback Neural Networks (RNN) in a new way, which are already accessible to billions of users today, for example for speech recognition on smartphones.

Biological brains are still superior to today's RNN in many ways. You will learn, among other things, a predictive world model that predicts how the environment will change as a result of the actions taken, and somehow use this world model for abstract thinking and planning. They continuously expand previously learned skills and thereby become more and more general problem solvers.

But we are working on the birth of a revolutionary RNN-based Artificial Intelligence (RNNAIssance) that can do that too.

An initially stupid agent converts the flow of perception from his environment through a control RNN called C into world-changing action sequences. His separate world model RNN called M always tries to discover regularities in the growing history of experiences, that is, to compress them, for example through predictive coding. In doing so, M always learns new parallel-sequential neural subroutines that represent what is predictable in compact form. C is rewarded for finding programs (or weight matrices) that solve user-specified problems. In order to reduce the search space, C can learn to navigate, awaken, question and otherwise exploit relevant parts of M in any calculable way in order to dramatically accelerate his search by "thinking".

In principle, it is clear how to do it. And it is not evident that conceptually something essential is still missing.

You say in lectures that before the end of this century the main decision-makers in this solar system will no longer be humans, but artificially intelligent systems. Sounds a bit like sci-fi.

Schmidhuber: Sounds like sci-fi, actually. But that's how it will go. Very cheap computers will soon be able to offer as much computing power as a human brain. And every ten years the computing power per euro increases by a factor of around 100. If the trend continues, we will only see cheap computers 50 years later with the combined computing power of all ten billion human brains.

And then the development will not stop, and there will be not just a few of these computers, but billions and billions. Not all of them will stay here in the biosphere. Most of them will be driven to where most of the resources are, i.e. out into space, where billions of self-replicating robot factories, gigantic telescopes and all sorts of things can be built in the asteroid belt in order to better understand how the universe works. Within a few million years, AIs will colonize the entire Milky Way, although humans will not play a significant role, contrary to the wishful thinking of many sci-fi films.

Human beings will not be able to control AI

Again to an earthly problem: a lot of people warn against artificial intelligence. They believe that at some point the development can no longer be controlled. Others think that's nonsense.

Schmidhuber: Many have seen Arnold Schwarzenegger films or "The Matrix" with ridiculous trade-offs between humans and the AI ​​systems of the future. These conflicts are usually completely freaking out.

In the long run, however, humans will indeed not be able to control their artificial creatures. These will no longer be mere tools. They will be far smarter than us, but at some point they will also lose interest in us, unlike in the films mentioned at the beginning.

Ultimately, you are always interested in those who are similar to you and with whom you share goals. You can only properly collaborate or argue with them. An extreme form of collaboration is love, an extreme form of competition is war. People are therefore interested in other people, artists in other artists, politicians in other politicians, five-year-old girls in other five-year-old girls. And the super-smart AI systems of the future will be interested in the super-smart other AI systems - and not so much in people.

I didn't want to focus on Hollywood films. Rather, it is experts from your field who warn of the developments that could result from self-learning intelligent systems. It's not difficult to imagine that in the future, systems will diagnose diseases and suggest treatment options ...

Schmidhuber: ... we already have that today ...

... exactly. And in the financial world we have systems that buy and sell securities in high-frequency trading at breakneck speed without people being able to understand it. It cannot be completely dismissed that the technical development of such intelligent systems can lead to uncontrollable situations.

Schmidhuber: Of course, one has to work to get such dangers under control.

But what you have to fear most is people who are as intelligent as you and have access to the 60 year old hydrogen bomb technology that can wipe out civilization without AI.

Would you dare to predict when there will be AI systems whose intelligence exceeds that of humans?

Schmidhuber: Personally, I would be amazed if this didn't happen in the next few decades. And if many AIs didn't pursue their own goals on a massive scale, most of which should be completely detached from human goals.

In principle, it is definitely desirable that AI systems, robots, etc., assist people in different situations in life - for example in the Elderly care or healthcare, in finance and insurance.

Schmidhuber: In fact, all of these developments already exist today. And of course there is already enormous commercial pressure to build friendly AI systems and thus make their users happier. Take the smartphone: Today it recognizes your language and your face and that of other users. It can make suggestions for dining out. It measures your heartbeat and can send warnings to you or a doctor if necessary.

AI systems are already operating - without many being aware of it - in a wide variety of areas of life. And in a way that most people approve of.

And privacy?

Schmidhuber: Of course, the question of privacy arises when a smartphone knows my heart rate exactly at different times of the day, etc. This problem is assessed very differently in different countries. People in Germany are much more skeptical than in the USA and Great Britain, which can also be a competitive disadvantage, as the field is then left to the others.

Last question: You keep hearing that Germany is lagging behind in terms of digitization. But there are many examples that German medium-sized companies are very innovative. You yourself cite the example of Ernst Dickmanns, who, decades before Google, namely in 1995, drove a self-driving Mercedes from Munich to Denmark and back on the autobahn. What do you think of the backwardness of Germans when it comes to digitization?

Schmidhuber: This is of course bullshit. Many of the methods used today are Microsoft, Google, Baidu, Samsung